1 / 10

HS06 – The history

HS06 – The history. Michele Michelotto. Before Y2K. HEP application were running on IBM mainframe, VAX and later on Unix workstation Unit to measure the CPU performances CERN Unit (CU), the power of an IBM 168 mainframe (circa 1980), proposed bt Eric McIntosh

laird
Download Presentation

HS06 – The history

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. HS06 – The history Michele Michelotto

  2. Before Y2K • HEP application were running on IBM mainframe, VAX and later on Unix workstation • Unit to measure the CPU performances • CERN Unit (CU), the power of an IBM 168 mainframe (circa 1980), proposed bt Eric McIntosh • VAX 11/780 was described a one-MIPS machine because it was equivalent to IBM System/360 that ran one Mega IPS • Sometimes called VUP (Vax Unit of Performance) • A CU was equivalent to 4 Mips

  3. SPEC • SPEC (Standard PerfomanceEvalutation Corporation) • The score express how many time the machine is faster than a reference machine • SPEC CPU 89 • SPEC CPU 92 • CINT92 and CFP92. No electronic records available • SPEC CPU 95 • CINT95 and CFP95 – SPECint_rate95 and SPECfp_rate95 also available • SPEC CPU 2000 • This version was widely used for the LHC experiments Computing TDR and at during the first year of the GRID.

  4. Problems with SI2K • You have a computer node and you want to know the throughput in SI2K. How can I know the get it • I measure it myself. I need to buy the SPEC suite, know how to run it, measure for each different processor • I look up on spec.org site.

  5. Lookup on SPEC site • I look on Spec.org the rating of a machine with the same processor. • I normalize the score to the clock of my node • I take that score and multiply for the number of processor in my node • I take that value and add on all the node of the same type • Problem: Should I take the base or the optimized value? Does it scale linearly? Should I consider the compiler used? The compiler switches? • I still have the problems of linearity • Some architectures had a very high score on SPEC that was not reproduces by our application

  6. I measure • I buy the SPEC suite. I learn to make it run on my software environment. I measure the SI2K int on a process and I multiply by the number of CPU in the machine. • More correct since I use the compiler I’ll use for my application but each site has its own set of version of compilers and compilations switches (CERN vsGridKa) • I still have the problems of linearity • I can’t use the rate because it is designed for job syncing at the end, not for batch jobs • Some architectures had a very high score on SPEC that was not reproduced by our application. But the effect with gcc is less evidents

  7. Using badly the SI2K • This confusion in the definition of SI2K was leading to a confusion in the definition of the CPU power of a center. • One can have two farm of 10 nodes each with 2 processor and 2 core each with SI2K scores different by a factor2 • If you overestimate your farm you can comply to the pledge but the your site can’t process all the HEP events you wanted to process • If you underestimate your farm you’re buying more CPU power than is necessary

  8. Requirements for HS06 • Match around 3% of main HEP application (simulation, reconstruction) • Footprint around 2GB/core like CPU application • Reproduce the batch job environment. Single thread programs on fully loaded worker node • Easy to run in order to have a number without any tuning • Easy to run for IT companies that participate to tender • Use the compiler switches from the architect forum

  9. Bonus • We made a fine tuning adding some Floating Point benchmarks taken from the Floating Point suite and added to the Integer Suite. • Otherwise the SPEC CPU INT 2006 would have been ok • Or a subset of it to keep it shorter

  10. Ageing • We imposed the –m32 switch to force 32 bit code on 64 bit since in that moment we were in early phases of the migration to 64bit • This is not a problem if the improvement of performances passing from 32 to 64 bit is the same on HS06 and HS14

More Related